TrueOpen
Get Started
← Back to Home

Verifying AI Inference

Verification must detect computation that violates the agreed rules while keeping costs under control. TrueOpen uses two verification approaches based on how large language models perform inference and on cryptographic proof techniques, reducing the cost of recomputation and proof generation.

Compute independently → Commit values → Reveal and compare

Logprob Majority-Consensus Verification

Verifiers feed the original input and the generated token sequence into the agreed model for a single parallel prefill pass.

Following the model’s rules, they check:

  • The generated tokens’ log probabilities and rankings.
  • The required Top-K candidate distributions.
  • Whether each output token is within the allowed candidate set.

These checks help detect model substitution or generation that violates the agreed rules.

For MoE models, Verifiers also compare the expert IDs selected at specified layers and positions. They do not need to regenerate the answer token by token.

Three Verifiers compute independently and commit their values before revealing them for comparison, preventing them from copying one another’s results.

Numerical tolerances and batch statistics reduce false judgments. Majority consensus requires valid, consistent results from at least two Verifiers.

Commit materials → Randomly select layers → Apply the model’s verification rules

Sampled Layerwise Cryptographic Verification

Cryptographic commitments bind the model and computation materials before random sampling, preventing later substitution. For fixed-point quantized models, proofs cover mandatory checks and sampled layers. For floating-point models, randomized matrix product checks determine whether numerical deviations fall within the agreed tolerances.

Sampling reduces verification costs, and its coverage determines the chance of detecting cheating. Stakes and penalties increase the losses that cheating nodes face when caught.

Explore the measurements behind the verification research and the related literature.

Verification cost

Qwen2.5-7B: 1.9%–14.8% of Generation GPU Time

Three prefill replays compared with generation time under the reported test conditions.

Read the experiment summary →Dense model identity

Qwen3-32B & 8B: Model Substitution Signals

Logprob comparisons across GPU hardware, quantization settings, and smaller-model substitutes.

Read the experiment →MoE model identity

Qwen3.6-35B-A3B: Logprobs & Expert Routing

Why logprobs alone are insufficient and how layer-0 routing helps distinguish model configurations.

Read the experiment →Preprint

Seal, Then Sample: Sampled Layerwise Proofs

The TrueOpen write-up of the SLP proof of concept, from GPT-2 to Llama-2-70B, on arXiv.

Read the preprint ↗Related literature

Model Auditing & Verifiable Computation

External papers, their relevance to TrueOpen, and the limits of that connection.

Browse the papers →